Anthropic, the AI company behind the Claude models, described how its Claude Code team reviews code changes. Before a pull request (PR, a proposed code change) opens, Claude runs tests and linters, which are automated error checks, and a separate agent reviews the change. More AI agents then hunt for bugs, others verify each finding, and a security pass can block the merge. Human approval depends on risk: none for low-risk changes, one reviewer for medium, two for high. The company says merged PRs more than doubled from June to September, while reverts fell by over 40%.
The key fix was a cap of three review rounds, which cut endless back-and-forth between agents. The company says PRs passing after one round rose from 65% to 78%, and those needing five or more rounds fell from 8.9% to 1.5%. Humans still own architecture and high-risk changes. The figures are internal and not independently verified.
